cifar10-models

ZoneTwelve CIFAR-10 model suite

This repository contains checkpoints from the MPS-trained CIFAR-10 model comparison suite. Each .pt file preserves model weights, optimizer state, epoch history, architecture metadata, and validation-selection metadata.

Public training-loss dashboard

Training loss for every architecture is synchronized to one public W&B project, with separate named runs (DenseCNN-Run-1, CNN-C-BN-Run-1, ViT-Run-1, and others). Each run includes raw loss and time-weighted EMA loss (decay=0.99).

Open the public W&B training-loss project ↗

CUDA reproduction

The companion CUDA model repository contains the NVIDIA A40 runs: GPU-resident tensor shards, GPU-side augmentation, batch size 1024, seed 42, and a 200-epoch target. Its selected results are ResNet-50 91.80%, DenseCNN 91.50%, CNN-C + BN 88.35%, and ViT 80.15%. The public CUDA W&B project contains every run and training-loss history. These are kept separate from the MPS reference results because the accelerator and input pipeline differ.

Fairness note: ResNet-50, DenseCNN, CNN A/B/C, ViT, and MobileNet-A now have 200-epoch results. R-CNN-B and MobileNet-B remain 30-epoch runs. Equal epoch count does not imply identical optimization recipes or compute.

DenseCNN attention maps

DenseCNN versus CNN Best

This comparison uses identical test images for DenseCNN (92.53%) and CNN Best, the reader-facing alias for CNN-C + BN (88.48%). Both models correctly classify the cat and ship, and both confuse the truck with an airplane. On the dog image, DenseCNN predicts cat at 98.4% while CNN-C + BN correctly predicts dog at 59.2%. Higher aggregate accuracy therefore does not mean winning every sample.

DenseCNN gallery

The gallery contains three correct predictions followed by three high-confidence mistakes: truck→airplane, dog→cat, and frog→bird. Each sample shows the input, heatmap, and overlay. Grad-CAM indicates sensitivity, not causal reasoning.

Top-10 contributing channels

Correct cat prediction:

Incorrect truck→airplane prediction:

Channels are ranked by the absolute gradient-weighted mean activation contribution to the predicted class. The truck error localizes the object but overweights its horizontal silhouette and horizon contrast, illustrating that correct localization can still produce the wrong semantic class.

Performance benchmarks

The following charts show MPS inference performance for every completed checkpoint benchmark. They measure execution cost only; they do not rank model accuracy.

Raw per-run JSON results and the combined summary are in the performance/ directory. Benchmarks used synthetic CIFAR inputs, batch size 128, 10 warmup iterations, and 50 measured iterations.

Final results

The full per-variant report is available in reports/ in the source repository.

Best / target epoch means best-validation checkpoint / final training target. For example, 192/200 means epoch 192 supplied the reported test checkpoint after a 200-epoch run.

* DenseCNN uses a different optimization recipe from the baseline models; matched epoch count is not identical compute.

Source code and synchronized reports are available at https://github.com/ZoneTwelve/cifar-baselines.